What is hallucination in generative AI?
A hallucination is a confident-looking model output that is unsupported, fabricated, or wrong.
- Looks fluent but is false
- RAG can reduce it but not eliminate it
- Verification still matters
Quick recall
374 cards
What is MCP in AI tooling?
MCP is a protocol pattern for connecting models to external tools, resources, and structured context in a standard way.
- Model talks through a protocol
- Tools and resources are exposed cleanly
- Goal is interoperability
What is model context window?
The context window is the maximum amount of tokenized input and working context a model can handle in one interaction.
- Includes prompt and retrieved context
- Longer is useful but not free
- Too much context can still hurt quality
What is model routing?
Model routing chooses different models or configurations based on task complexity, latency needs, cost, or safety requirements.
- Cheaper models for simple tasks
- Stronger models for harder tasks
- Routing policy is a product decision
What is prompt engineering?
Prompt engineering is the practice of structuring instructions and context so a model produces more reliable outputs.
- Instruction quality matters
- Examples can shape behavior
- Evaluation is still required
What is prompt injection?
Prompt injection is when untrusted input tries to override instructions or manipulate the model into unsafe behavior.
- Often comes through retrieved or user content
- Model sees it as text not trusted code
- Needs system design defenses
What is tool calling in LLM apps?
Tool calling lets a model request an external function or service when the answer requires real actions or fresh data.
- Model decides a tool is needed
- App executes the tool
- Model continues with the result
Why do AI applications need human-in-the-loop workflows?
Human-in-the-loop design adds review or approval where model mistakes would be costly, risky, or hard to reverse.
- Use for high-impact actions
- Humans handle ambiguity and accountability
- Not every flow needs it
Bias vs variance
Bias is error from overly simple assumptions; variance is sensitivity to training data noise.
- High bias misses pattern
- High variance chases noise
- Need a balance
Classification vs regression
Classification predicts categories; regression predicts continuous numeric values.
- Class labels vs numbers
- Metrics differ
- Thresholding often matters for classification
Model monitoring basics
Model monitoring tracks data drift, prediction quality, latency, and operational health after deployment.
- Watch for drift
- Track business metrics too
- Retrain only when signals justify it
Overfitting vs underfitting
Overfitting memorizes noise; underfitting is too simple to capture the real pattern.
- Overfit has low train and poor test performance
- Underfit performs poorly everywhere
- Control with data, features, and regularization